xref: /aosp_15_r20/external/executorch/backends/qualcomm/builders/op_quantize.py (revision 523fa7a60841cd1ecfb9cc4201f1ca8b03ed023a) 
1# Copyright (c) Qualcomm Innovation Center, Inc.
2# All rights reserved
3#
4# This source code is licensed under the BSD-style license found in the
5# LICENSE file in the root directory of this source tree.
6from typing import Dict
7
8import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
9
10import torch
11from executorch.backends.qualcomm.utils.constants import QCOM_ENCODING, QCOM_QUANT_ATTRS
12
13from .node_visitor import NodeVisitor, register_node_visitor
14from .qnn_constants import OpQuantize, QNN_OP_PACKAGE_NAME_QTI_AISW
15
16
17class QuantizeOpBase(NodeVisitor):
18    def __init__(self, *args) -> None:
19        super().__init__(*args)
20
21    def define_node(
22        self,
23        node: torch.fx.Node,
24        nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
25    ) -> PyQnnWrapper.PyQnnOpWrapper:
26        quant_input_tensors = []
27        input_node = node.args[0]
28        input_tensor = self.get_tensor(input_node, node)
29        inp_tensor_wrapper = self.define_tensor(
30            input_node,
31            input_tensor,
32            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
33            nodes_to_wrappers,
34            is_input_tensor=True,
35        )
36        quant_input_tensors.append(inp_tensor_wrapper)
37
38        node.meta[QCOM_QUANT_ATTRS] = {QCOM_ENCODING: node.target}
39        arg_schemas = list(node.target._schema.arguments)[1:]
40        for i, arg_schema in enumerate(arg_schemas):
41            name = arg_schema.name
42            node.meta[QCOM_QUANT_ATTRS][name] = node.args[i + 1]
43
44        output_tensor = self.get_tensor(node, node)
45        output_tensor_wrapper = self.define_tensor(
46            node,
47            output_tensor,
48            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
49            nodes_to_wrappers,
50            is_input_tensor=False,
51        )
52        quant_output_tensors = [output_tensor_wrapper]
53
54        quant_op = PyQnnWrapper.PyQnnOpWrapper(
55            node.target.__name__,
56            QNN_OP_PACKAGE_NAME_QTI_AISW,
57            OpQuantize.op_name,
58        )
59        quant_op.AddInputTensors(quant_input_tensors)
60        quant_op.AddOutputTensors(quant_output_tensors)
61
62        return quant_op
63
64
65@register_node_visitor
66class PerTensorQuantize(QuantizeOpBase):
67    target = ["quantized_decomposed.quantize_per_tensor.default"]
68
69
70@register_node_visitor
71class PerChannelQuantize(QuantizeOpBase):
72    target = ["quantized_decomposed.quantize_per_channel.default"]
73